About

Why Label3D?

Label3D was designed to streamline the generation of training sets for multi-view 3D pose estimation and tracking algorithms in scientific applications. It addresses three problems that typically arise when generating 3D keypoint estimates from multi-view images without continuous triangulation.

1. It is difficult to identify inaccurate calibration across cameras when labeling images serially.
2. It is difficult for humans to accurately label the exact points in images that correspond to a single point in 3D.
3. It is time consuming to label points across multiple images.

How does it work?

Label3D addresses these problems by simultaneously displaying multi-view images and requiring continuous triangulation of labeled points. In the order above:

1. Inaccurate calibration is immediately identifiable when triangulated points do not align with one another.
2. Labels across all images are by definition in agreement with a single point in 3D.
3. One can confidently label a point across any number of views with as few as two perspectives.

Other

In addition to the technical benefits above, Label3D offers a number of practical benefits.

• Easy multi-view reprojection of 3D pose estimates
• Easy active labeling
• Easy debugging for camera calibration and frame synchronization
• Easy video writing through Animator superclass
• Integration with Animator classes for custom interactive data visualizations

For the full list of features, please see the Documentation page.